Deterministic dynamics of distributional multi-agent reinforcement learning.

Distributional multi-agent reinforcement learning exhibits choice hysteresis and perseveration driven by return discretization, revealing new insights into decision-making dynamics.

• Why it matters: Understanding how cognitive biases like optimism influence behavior across different contexts is crucial for bridging neuroscience, psychology, and social sciences, yet remains fragmented.
• What they did: The authors developed a deterministic framework that models the dynamics of distributional reinforcement learning using a finite set of neurons to discretize return distributions, validated across three key domains.
• The result: This approach uncovers how return discretization can cause stable yet suboptimal strategies and offers mechanistic explanations for incoherent choices, advancing hypotheses on cognitive bias propagation in complex social environments.